The said hair product is probably Pedro L'Âne

Already have my train tickets and hotel booked, very much hope that the HPC devroom will be there!

For some time there was the argument that Linux was no good for gaming, but in the past few years I heard more and more good things about gaming on Linux so it's not anymore an argument.

To add some weight for Linux: most of the open source software is more than good enough (LibreOffice, Gimp, Inkscape, VLC...) and for personal use I've never needed anything more.

Linux is obviously a good choice, it's lightweight, it mostly just works, you have full control of your computer and even for non technical stuff, it's much easier to do things in bash than with a GUI (a good example is changing metadata from thousands of pictures)

From the little experience I have with Mac I find them frustrating because it looks like Linux but it's not with slight differences that makes you bang your head against the wall sometimes. Probably less frustrating for *BSD users though. Oh and the Mac keyboard is very confusing to me too.

Only Linux for the past 18 years (work and personal use), mostly Dell and ThinkPad. No real issues apart from the occasional driver problem on Dell. On the long run ThinkPads are a bit more solid, but Dell are pretty good (the non pro versions less so). Don't like the ThinkPad keyboard so much.

I was writing exactly the same; there's an "infrastructure" side of putting up such a cluster that is not visible with a laptop. If you are an end-user, a couple of containers is great. If you want to understand how a cluster works with its constraints, then a RPi cluster is great.

There's value in doing so for educational purposes though, ICHEC has a little cluster they displayed at events like ISC, and I remember a talk at FOSDEM (archive.fosdem.org/2020/schedul...) about it.

Having the flops doesn't mean you can use them with your code. Using them doesn't mean your algo & science is solid. Using them w/ solid algo/physics doesn't mean you can efficiently analyze the data. Analyzing the data doesn't mean you have a profound understanding of the problem you study...

And there are already real-life examples where the efficiency of AI exceeds by far the classical approach one, for both power and accuracy. One of my favourite examples for the past couple of years has been FourCastNet (docs.nvidia.com/deeplearning...); this kind of achievement is amazing

It's not that surprising as -- at least in Europe -- lots of AI factories are actually pre-exascale and exascale EuroHPC systems.

And being maybe too optimistic and excited about the future, I'd like to gamble on quantum computing. But more realistically that might be the next-next 10x 😁

Focusing on raw performance of apps, I'd go for AI impact: there is a tremendous potential with PINNs and the likes, and the way it revolutionised weather forecasts for example is just the beginning. And there's also more than promising things with accurate subgrid modeling with AI.

I'd go with better usability, because if you build big systems but don't care about services you provide for the users to make the best of it, you've lost. It's kind of the human factor in the equation. Incidentally, I/O & SW stack have an impact on usability.

Quite some work was done on the doc to ease the use of the library and broaden its user base.
You want to run containers at scale with no hassle? You have a legacy code & want to switch MPI? You want to debug or assess perf easily? Wi4MPI might be for you!

wi4mpi.readthedocs.io/en/latest/us...